This paper introduces an end‑to‑end framework for building telecom language models that combine broad networking knowledge with specialist behavior in routing, broadband services, access technologies, and operational diagnostics, while remaining practical to deploy as a single model. Synthetic, source‑grounded training data are generated from structured networking taxonomies and expert graphs, yielding 212,559 routed records across fourteen networking routes. Six frozen LoRA specialist teachers are trained on a shared continued‑pretrained backbone, each covering a specific route.
The core distillation method, Routed Specialist On‑Policy Distillation (RSOD), works as follows: the student generates each trajectory, the route‑assigned specialist teacher forces the exact native token sequence, compact teacher hidden states are reconstructed via a shared frozen output head, and only the student LoRA is updated. Controlled reconstruction validation reports a mean absolute logit error of $7.26\times10^{-6}$ with 100% top‑1 agreement between native and reconstructed teacher logits.
Training on four RTX 6000 Ada GPUs for 2,000 steps produced 16,000 trajectories with zero observed student‑version or token‑ID mismatches. In the primary route‑conditioned evaluation on 2,850 held‑out questions, the raw Meta‑Llama‑3.1‑8B‑Instruct model scores $12.8661/25$ (micro) and $13.0172/25$ (macro); the combined six‑teacher routed system scores $21.7641/25$ (micro); the consolidated RSOD student, released as Llama‑CalixCLM v1, reaches $21.7112/25$ (micro) and $21.8006/25$ (macro). Under a neutral deployment‑like prompt, the same student attains $20.0028/25$ (micro) and $20.0778/25$ (macro). The student remains within $0.0529$ micro‑score points of the routed‑teacher system while preserving a numerically validated and auditable train‑serve execution path.
Review: RSOD successfully merges multi‑expert capabilities into a compact, deployable model with negligible error and performance virtually indistinguishable from the full routed‑teacher ensemble, offering a promising direction for telecom‑focused large‑model distillation.